Agnes 3.0 Flash: Apache-2.0 33B multimodal that fits one H100
Sapiens AI ships open weights scoring 85 on GPQA, with text-image-video input and OpenAI-compatible serving, plus a sober result on self-tuning agents.
Copy markdown
Frontier-grade reasoning that fits one GPU
Agnes 3.0 Flash is a 33B open-weights model under Apache 2.0 — roughly 66GB of weights that load on a single H200 (141GB) or H100 (80GB) with 128GB+ host RAM. You can self-host it and ship commercially, with no API contract in the way.
85 on GPQA Diamond, with an asterisk
Reported scores land at 85.05 GPQA Diamond and 74.20 IFBench, with a reasoning-effort dial (high/medium/low/off) to trade tokens for accuracy. The numbers are self-reported and it trails on agentic and terminal tasks, so benchmark it before you bet a pipeline on it.
Text, image, and video in, tool calls out
It reads text, images, and video across a 262K-token context and emits tool calls via <tool_call>, served through Transformers or SGLang on OpenAI-compatible endpoints. Build this weekend: a self-hosted agent that reads screenshots or video frames without renting a frontier API.
Or rent it for almost nothing
Not ready to run a GPU? Hosted endpoints list at $0.05 per million input tokens and $0.15 output, with a 90% cache discount — cheap enough to prototype against before you commit to your own hardware.
Elsewhere: don't let an agent rewrite its own harness
ByteDance Seed's new HarnessDev benchmark let LLMs engineer their own agent scaffolding, and only 34 of 64 changes generalized to held-out tasks. If you auto-tune your agent's harness, validate on tasks it never saw — otherwise you are just overfitting.